Conceptual design of multidimensional cubes with LLMs: An investigation

Abstract

Large Language Models (LLMs) can simulate human linguistic capabilities, thus producing a disruptive impact across several domains, including software engineering. In this paper we focus on a specific scenario of software engineering, that of conceptual design of multidimensional data cubes. The goal is to evaluate the performance of LLMs (precisely, of ChatGPT-4o) in multidimensional conceptual design using the Dimensional Fact Model as a reference. To this end, we formulate nine research questions to (i) understand the competences of ChatGPT in multidimensional conceptual design, following either a supply- or a demand-driven approach, and (ii) investigate to what extent they can be improved via prompt engineering. After describing the research process in terms of base criteria, technological setting, input/output format, prompt templates, test cases, and metrics for evaluating the results, we discuss the output of the experiment. Our main conclusions are that (i) when prompts are enhanced with detailed procedural instructions and examples, the results produced significantly improve in all cases; and (ii) overall, ChatGPT is better at demand-driven design than at supply-driven design.